基于XGBoost算法的烤烟采收成熟度图像识别

Image recognition of flue-cured tobacco harvest maturity based on XGBoost algorithm

【目的】为实现智能精准识别烟叶采收成熟度。【方法】以云烟87为试验材料,利用OpenCV和灰度共生矩阵(GLCM)提取图像特征,构建极限梯度提升(XGBoost)算法模型从而实现对鲜烟叶成熟度识别。【结果】①鲜烟叶图像特征中,R(红,red)、G(绿,green)、B(蓝,blue)分量和ASM(角二阶矩)随着成熟度的增加呈现较为明显的上升趋势,其他图像特征变化不显著;②经F分数(F-score)、AUC值(受试者工作特征曲线与坐标轴之间的面积)和准确率逐步筛选,得出R1(R分量均值)、G1(G分量均值)、B1(B分量均值)、S2(S分量方差)和B2(B分量方差)等5个特征参数,据此建立的XGBoost算法模型对烟叶成熟度识别准确率达到95.85%,比22维特征参数建模的准确率高0.41%,比BP神经网络模型高4.71%。【结论】基于机器视觉下的XGBoost算法可准确、高效地识别鲜烟叶成熟度。

[Objective]This study aims to achieve intelligent and accurate identification of tobacco harvest maturity.[Methods]Taking Yunyan 87 as the test material,we extracted image features using OpenCV and GLCM,and constructed XGBoost algorithm model so as to realize the maturity recognition of fresh tobacco leaves.[Results]①In the image features of fresh tobacco leaves,the components R(red),G(green),B(blue),and ASM(Angular Second Moment)showed a significant rising trend with the increase of maturity,while other image features did not change significantly;②After stepwise screening of F-score,AUC value(Area Under the Receiver Operating Characteristic Curve),and accuracy rate,five feature parameters including R1(mean value of R component),G1(mean value of G component),B1(mean value of B component),S2(variance of S component),and B2(variance of B component)were selected.The XGBoost algorithm model established based on these features achieved an accuracy rate of 95.85%in identifying tobacco leaf maturity,which is 0.41%higher than the model with 22-dimensional feature parameters and 2.72%higher than the BP neural network model.[Conclusion]The XGBoost algorithm based on machine vision can accurately and efficiently identify the maturity of fresh tobacco leaves.